Predictive calibration for digital sun sensors using sparse submanifold convolutional neural networks

Recent developments in AI for space applications mirror the success achieved in terrestrial domains, particularly for data-rich computer vision tasks. Among spacecraft attitude sensors, digital sun sensors (DSS) are widely used due to their simplicity, robustness, and low cost, but achievable attitude accuracy is often limited by sensor errors and calibration quality. While these errors can be mitigated through calibration, traditional DSS calibration procedures are costly, time-consuming, and labor-intensive, and existing calibration algorithms rely on shallow features such as centroids, are not end-to-end, require segmentation, are mask-specific, and noise-limited. In addition, conventional multiplexing approaches typically rely on explicit template matching for correlation mapping. This work focuses on predictive calibration for improving DSS performance, motivated by practical constraints relevant to nanosatellite platforms, where limited hardware tuning and manufacturing precision increase reliance on high-quality calibration. To address the limitations of conventional calibration approaches, we propose an end-to-end predictive calibration framework for DSS using a sparse submanifold convolutional neural network (SSCNN) to estimate two-axis sun angles directly from illumination images. The proposed method achieves sub-arcminute accuracy, approximately 0.005°mean absolute error, for the evaluated mask configurations and within the studied local FOV region bounds. Following a proof-of-principle evaluation, this work focuses on the local mapping of four FOV regions via global normalization to demonstrate network prototyping and evaluation, with local normalization introduced to support scalable FOV expansion for future system-level scaling exercises. The network is used solely for method prototyping and pre-training and is not intended to be flight-ready; bridging this Sim2Real gap is identified as future work, with domain adaptation recommended for real-space deployment. The main contributions of this work include the application of a CNN-based regression framework to DSS predictive calibration, an end-to-end calibration approach that replaces traditional centroid-based pipelines, a multiplexing-capable network in which local spot features and correlation rules are learned representations rather than an a priori coding template, and the development of a publicly available physics-based synthetic DSS dataset and simulation framework.